What the data for this category looks like
Core data for ad creative review originates from ad delivery platforms, landing page resources, product qualification documents, and general advertising law compliance clauses. Daily processed documents include single ad creative copy, landing page screenshots, historical review opinions, and creative modification records. Data updates occur daily or in real time, and synchronize with ad delivery plan adjustments or compliance clause updates. Single review tasks associate with diverse document types. Three core information categories must be covered: delivered content, compliance basis, and historical processing results. This scenario is compatible with FastGPT v4.17.0.
Constraints imposed on knowledge base retrieval and recall
Multi-source data sources require retrieval systems to support unified indexing across document formats. This prevents missing compliance clauses or review record recalls due to document type differences. Real-time or daily data updates require knowledge bases to support incremental synchronization. This ensures the latest compliance clauses and creative modification records can be retrieved quickly. Diverse document types require retrieval logic to balance semantic association and exact keyword matching. It must identify the semantic intent of non-compliant copy, and match modification requirements in specific compliance clauses or review opinions. High-risk error outcomes require retrieval results to have high confidence. The proportion of irrelevant content in recalls must be strictly limited, to avoid misleading review decisions.
Configuration settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Ad creative review requires precise matching of compliance clauses and non-compliant content. This range filters low-relevance irrelevant results |
limit | Top 8 entries | The context length of a single review task must be controlled within a reasonable range, to avoid exceeding the model's processing limit |
searchMode | Semantic retrieval + keyword hybrid | Balances the need for semantic intent matching of ad copy and exact keyword matching of compliance clauses |
usingReRank | Enabled | Performs secondary ranking on initial recall results, prioritizing content directly related to violation point annotations and compliance basis |
datasetSearchUsingExtensionQuery | Enabled | Expands reviewer questions into compliance-related query terms, improving the recall accuracy of historical review records |
collectionIds | Bind dedicated collection IDs for ad creative review | Isolates knowledge bases across different business scenarios, preventing cross-category review data from interfering with retrieval results |
Three common configuration mistakes
- Incorrectly filling non-dedicated collection IDs into the
collectionIdsconfiguration item. The symptom is that retrieval results mix document content from account opening review or other businesses. The cause is failure to isolate knowledge base collections by business scenario. - Not configuring the
similaritythreshold or setting the threshold too low. The symptom is that queries unrelated to ad compliance still trigger knowledge base retrieval and return relevant results. The cause is failure to set a minimum relevance limit to filter irrelevant content. - Not enabling
datasetSearchUsingExtensionQueryand not adjusting content priority rules. The symptom is auxiliary data such as modification records in historical review opinions are not prioritized for recall. The cause is failure to expand query scope to cover associated information of auxiliary documents.
How to verify correct configuration
- Upload dedicated documents such as ad creatives and review opinions. Check if they are automatically categorized into the configured dedicated collection. Confirm that
collectionIdsis bound correctly. - Input test questions unrelated to ad compliance. Verify if preset non-knowledge base responses are triggered. Confirm that
similarityrelated restriction configurations take effect. - Input review questions containing non-compliant keywords. Check if recall results prioritize compliance clauses and historical review records. Confirm that
usingReRankand retrieval mode configurations take effect. - View the number of retrieval return results. Confirm that they meet the quantity requirements of the
limitconfiguration. Verify that the retrieval upper limit setting is correct.
The values provided are common starting points, and should be evaluated against the samples relevant to the specific use case.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.